Fourier Transform

Images are samples of different intensities that can be represented in the frequency domain. Frequency representations enable powerful analysis and manipulation techniques.

Lesson Objectives

  • Using sines and cosines to reconstruct a signal

  • The Fourier Transform and its properties

  • Frequency domain representation of signals and images

  • Three properties of convolution (commutativity, associativity, distributivity)

Fourier Transform

The Fourier Transform decomposes a signal (or image) into a sum of sinusoidal components at different frequencies, amplitudes, and phases. This reveals which frequencies dominate the signal.

  • Low frequencies correspond to smooth, slowly varying regions

  • High frequencies correspond to edges and fine detail

  • Filtering in the frequency domain (e.g., low-pass, high-pass) is equivalent to convolution in the spatial domain

Blending

Frequency-domain analysis informs blending strategies — e.g., blending low frequencies from one image with high frequencies from another (hybrid images).

Pyramids

Gaussian and Laplacian pyramids provide a multi-scale frequency decomposition useful for blending, compression, and analysis.

Cuts

Graph cuts and min-cut/max-flow algorithms find optimal seams for compositing images.

Features

Feature detection identifies distinctive points in images for matching, alignment, and recognition.

Feature Detection and Matching

Techniques for detecting and matching features across images:

  • Harris corner detector: Finds points with large intensity variation in multiple directions

  • SIFT (Scale-Invariant Feature Transform): Detects and describes features invariant to scale and rotation

  • Feature matching: Comparing descriptors between images to find correspondences